Neural Networks versus Genetic Algorithms as Medical Classifiers - Foundations on Natural and Artificial Computation - page 398

Finally, the results obtained after performing multiple training and testing
tests using the genetic algorithm classifier and the neural networks based, have
been the ones we see in(Fig. 4):
Fig. 4. Obtained accuracy rates
4 Conclusions
The purpose of our study was to test the feasibility of using GAs as classifiers
in a medical enviroment. Since there are other soft computing tools that have
been widely tested for this purpose, as it is the case of ANNs, the comparison of
the diagnostic results obtained using a GA implementatio and the results given
by these other well-prooved classifiers, can be interpreted as a measure of how
good the diagnostic performance of a GA implementation could be.
In our experimentation, we have designed and implemented three classifiers
based on two different techniques from machine learning field. The objective
was to implement an automatic system for the diagnosis of two urinary tract
dysfunctions: urinary incontinence and urinary tract obstruction. To do this, we
have collected data from a clinical database that contains samples from real test
performed over patients diagnosed from suffering one of the two anomalies.
Two of the classifiers are based on two supervised artificial neural network
models, a multi-layer perceptron and a radial basis function neural network.
The third classifier was implemented using a genetic algorithm. There have been
performed some tests using different parameter's configurations of each of them,
to get one that provided a better accuracy rate in the diagnostic process.
Analysing the accuracy rates we can see that they were quite good in general
terms of accuracy. The classifier based on a MLP gave the best results but
closely followed by the one based on GA. On the other hand, despite offering
good diagnostic rates, the RBF based classifier is far from being as accurate as
the other classifiers.
These results show that GAs are as good as ANNs in clinical diagnostic tasks.
This fact is a first step in choosing an optimal machine learning technique in
order to implement a more broad and general system for diagnosing that allows
us to discriminate between healthy patients and those presenting any anomaly
on the urinary tract.
Future projects based on the results of this study are getting a larger number
of clinical tests samples, to have a greater clinical database that could cover
a wider range of urological anomalies, to test new types of machine learning
techniques such as unsupervised neural networks, competitive, etc.